cross-modal_patch_attack's Introduction

Unified Adversarial Patch for Cross-modal Attacks in the Physical World

This is an official implementation 🍎 and we aim to implement an effective rgb-infrared multi-modal attack in the physical world.

🍊[Read our arXiv Paper]

🍇 Usage

Install

Python>=3.6.0 is required with all requirements.txt installed including PyTorch>=1.7:

$ git clone https://github.com/Aries-iai/Cross-modal_Patch_Attack
$ cd Cross-modal_Patch_Attack
$ pip install -r requirements.txt

Data Convention

The data is organized as follows:

dataset  
|-- attack_infrared
    |-- 000.png        # images in the infrared modality
    |-- 001.png
    ...
|-- attack_visible
    |-- 000.png        # images in the visible modality
    |-- 001.png
    ...

Here, we should ensure the consistency of infrared images and visible images' names.

Running

python spline_DE_attack.py

Notes

When you prepare dataset, you need to change directory path in spline_DE_attack.py and DE.py.
⚠️ If you want to attack other detection models, you need to change yolov3 folder to the model folder you want to attack and add detect_infrared.py, detect_visible.py in this new folder for returning targets' detection confidence scores to DE.
The weights of yolov3 models can be downloaded from: https://drive.google.com/file/d/1gpPnHcGRjrJAComQety__dWVwJTWCnTk/view?usp=drive_link.
The part of attacked images can be downloaded from: https://drive.google.com/file/d/1C7mhrr94lXu4qw_P1dX5hwpRuDc4iI-4/view?usp=drive_link.

🏫 Citation

Cite as below if you find this repository is helpful to your project:

@misc{wei2023unified,
      title={Unified Adversarial Patch for Cross-modal Attacks in the Physical World}, 
      author={Xingxing Wei and Yao Huang and Yitong Sun and Jie Yu},
      year={2023},
      eprint={2307.07859},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}